The Future Arrives in 2045
theconnectivist.com
theconnectivist.com
A current deep learning neural network cluster and "Hard AI" seem similar but they really bear no relation to each other. Like comparing a bird to an airplane, they both fly and have wings but the ability to make one isn't related to making the other. Right now we're building better birds, true AI is a stealth bomber.
People say that whenever computers achieve a goal, then the goal is no longer considered AI. For example people 20 years ago didn't think that computers could play chess or compete on Jeopardy, but now that they've done those things they aren't thought of as impressive demonstrations of intelligence any more. There's some truth to this, but for the majority of people the goalposts have never moved. They associate the term "artificial intelligence" with an artificial mind that functions in the same way that a human mind does, "Hard AI" in the tradition of Asimov and other science fiction writers. We're as far away from that as we've ever been. It could happen in 2045 or 4045, or anywhere in between. No evidence exists that we're getting closer and there's no reasonable way to predict when what we can't imagine will become reality.
I think the easiest method to achieve it, would be to first upload an actual human brain to a computer environment - by slicing a brain into molecule-thin slices, then observing the neural responses of that simulated brain.
The examples of "AI" cited in the article are remarkable, but are still extremely specific or not really intelligence at all. Siri, etc, are nothing more than text parsers that give a canned set of responses. The work on neural networks is interesting but still, at best, only a small component of actual AI. (Note: I'm not going to define an "actual AI". Yes, I know we keep moving the goalposts on what that would be. I'll know it when I see it, and so will you).
I'm not saying it won't happen, but it will require a type of conceptual breakthrough that we simply haven't had yet. To hype "the singularity is nigh!" at this point is dishonest, trivializes the real problems and sets false expectations for industry and policy-makers.
His idea is that progress is exponential in all of the requisite areas. That includes algorithms, hardware, biology, neuroscience, and more.
> Siri, etc, are nothing more than text parsers that give a canned set of responses
We say the same about everything once we can do it with computers, because progress doesn't come magically. It's incremental. Yet much of what we have already is what used to be "science fiction" and, before that, "magic". I'm sure when AIs are passing the Turing test, we won't think any more of computers, but less of the Turing test.
This isn't an argument against Kurzweil's predictions, it's just moving the goalposts as you say. If strong AI comes, we'll move them all the way there and maybe even a bit past.
I disagree that we'll know it when we see it though. I think we'll deny it until we die and the new generation grows up in a world in which computers have rights.
> I'm not saying it won't happen, but it will require a type of conceptual breakthrough that we simply haven't had yet.
Finally, Kurzweil's arguments do not require a breakthrough as a premise. Kurzweil's idea is to scan the brain at the neuronal level, then brute-force its simulation at the biochemical level. It's straightforward extrapolation.
You could propose that we will come across some roadblock in our exponential progress towards that goal, but in the absence of one, the null hypothesis is that progress will continue as it has. Then indeed, the singularity is nigh.
What I'm skeptical about are the predictions that we will reach something more intelligent than humans (how to quantify intelligence even?), that it will improve our culture, and other sci-fi stuff...
Even if it turns out the predictions are wrong, the brute force can reach something interesting anyway, not human-like intelligence, but something new and complementary.
I'd argue, however, that since the rate of change is accelerating, maybe we should actually compare to 60 years ago. In 1954 we were practically in the stone ages.
Intelligent things don't always do what they're told; that's what makes them intelligent. If you told someone to jump off a cliff, and they immediately did it, would you think "wow, that person was very intelligent"? Of course not. What if you tried to push them? I would expect an intelligent person to fight back.
Now replace that person with a robot. Do we really want a robot that will refuse to follow our commands? Do we really want a robot that will fight back against us? Even Asimov put self-preservation #3 in the list of rules, after following human commands. But I challenge you to think of an intelligent being that is not dangerous in some way, when threatened. I propose that this attribute is not separable from intelligence.
It seems to me that it is impossible to conceive of a truly intelligent artificial being without considering it dangerous. Bumblebees are dangerous; dogs are dangerous; people are certainly dangerous. But who is working on creating robots that are designed from day one to be dangerous to humans? I can't remember ever hearing of such a research program.
And I don't think that just "happens" when algorithms get complex enough. Not when the algorithms and even hardware are designed and built from an inherent assumption of obedience and compliance.
This article seems overly optimistic.